Neural virtual sensors for terrain adaptation of walking machines

Abstract

When walking in realistic conditions, accurate, reliable sensorial information is critical to ensure the safe operation of legged robots. That means a large number of sensors, cabling, and electronic systems must be used, complicating the robot. On the other hand, the great complexity of the hardware of walking machines is one of the main obstacles preventing the introduction of this kind of vehicle in real applications; consequently this hardware should be simplified. These antagonistic requirements can be reconciled by the use of what are called virtual sensors. This paper addresses the design of virtual sensors for terrain adaptation developed with the aims of simplifying the hardware of the walking machine or increasing the reliability of the sensorial information available. These virtual sensors are based on neural networks and can estimate the forces exerted by the feet from data extracted from joint-position sensors, which are mandatory in all robotic systems. The force estimates are used to detect foot/ground contact. Some experiments carried out with the SILO4 walking robot are reported to prove the efficacy of this method. © 2005 Wiley Periodicals, Inc.